# Bulletin of the Seismological Society of America

## Abstract

A local magnitude (*M*_{L}) relation for the western Canada sedimentary basin (WCSB) is developed using a rich ground‐motion dataset compiled from local and regional networks in the area. The assessment of amplitude decay with distance suggests that joining direct waves with postcritical reflections from Moho discontinuity modifies the attenuation pattern in the 100–200 km distance range. The *M*_{L} distance correction is parameterized using a trilinear function that accounts for the observed attenuation attributes. Regression of ground‐motion amplitudes results in the following distance correction model (−log*A*_{0}) for earthquakes in WCSB: in which *R* is the hypocentral distance (km). The standard *M*_{L} relations fail to capture the rates and shape of amplitude attenuation in the region, resulting in overestimated magnitudes by 0.3–0.6 units. The overestimation is larger for local networks due to the increased discrepancy between standard *M*_{L} relations and actual attenuation properties at close distances. The derived relationship results in unbiased *M*_{L} magnitude estimates in WCSB over a wide distance range (2–600 km), which ensures consistent magnitude estimates from local and regional networks.

## Introduction

Local magnitude (*M*_{L}) is the most widely reported magnitude scale in earthquake catalogs. It is primarily used as an initial estimate of earthquake size when quick information on seismic events is released to the public (Kanamori, 1983). Because of its ease of calculation in real time, *M*_{L} is commonly adopted in seismic monitoring applications as a driver of operational protocols designed to mitigate induced seismicity risks (e.g., Alberta Energy Regulator [AER], 2015). Local magnitude is also used for the estimation of moment magnitude (*M*_{w}) for small events where signal‐to‐noise ratios (SNRs) are insufficient for a moment tensor inversion (e.g., Ristau *et al.*, 2003, 2005; Bethmann *et al.*, 2011). These estimates are employed in derivation of ground‐motion prediction equations (GMPEs) and magnitude–recurrence relations for seismic‐hazard analysis.

Richter (1935, 1958) defined *M*_{L} based on the peak horizontal amplitude that would be recorded on a standard Wood–Anderson (WA) seismometer located at 100 km from an event. He derived a distance correction curve that projects the recorded amplitudes to the reference distance using ground motions from southern California. A proper projection of recorded ground motions to the reference distance (100 km) is the prerequisite for accurate *M*_{L} calculations. Therefore, a regional calibration to Richter’s original distance correction (−log*A*_{0}) is often required to consistently reflect the attenuation attributes in tectonic environments that differ from southern California (Di Bona, 2016). However, the standard magnitude relations (e.g., Richter, 1935, 1958; Hutton and Boore, 1987; Eaton, 1992) are generally used in areas where ground‐motion data are insufficient to examine regional attenuation effects. This may result in biased *M*_{L} estimates if the adopted magnitude relation does not comply with the attenuation characteristics of the target region.

In western Canada, *M*_{L} magnitudes reported by Geological Survey of Canada (GSC) are based on the original formula by Richter (1935). A preliminary assessment of Richter’s relation for a well‐recorded earthquake in western Alberta revealed inconsistent *M*_{L} estimates at different distances. Motivated by this observation, in this study, I examine the amplitude attenuation in the western Canada sedimentary basin (WCSB). A comparison of peak amplitudes from earthquakes and mining blasts shows that both event types decay at similar rates with distance. Both event types demonstrate a Moho‐bounce effect, a consequence of reflected and refracted phases joining the direct waves, between 100 and 220 km. I develop a calibrated *M*_{L} formula for WCSB by merging the two amplitude datasets. I characterize the regional −log*A*_{0} model using a trilinear function to account for the observed shape of attenuation. Additionally, a correction factor for each station is determined, relative to the average site condition in the region. This allows for consistent *M*_{L} estimations across different stations, regardless of their site conditions. I show that standard *M*_{L} relations fail to capture the rates and shape of attenuation in WCSB. This results in overestimation of event magnitudes by 0.3–0.6 *M*_{L} units, on average, depending on the adopted relation. The overestimation is larger for *M*_{L}<2.0 events, which are primarily detected by local networks, due to an increased discrepancy between standard *M*_{L} models and the actual attenuation attributes at close distances. The calibrated magnitude relation provides unbiased *M*_{L} estimates in WCSB, over a wide distance range from 2 to 600 km.

## A Preliminary Assessment on the Suitability of Richter *M*_{L} Formula for WCSB

GSC uses the original formula by Richter (1935) to compute *M*_{L} magnitudes of earthquakes in western Canada. Vertical‐component motions obtained at distances between 50 km<*R*<600 km are used in *M*_{L} calculations (Ristau *et al.*, 2003, 2005). Stations located at *R*<50 km are excluded to reduce the effect of focal depth. This also limits biased magnitude estimates that may arise from a potential discrepancy in close‐distance attenuation between western Canada and southern California. However, omitting close‐distance high SNR records is inconvenient for small events because it makes their *M*_{L} estimates more susceptible to noise effects. Besides, the suitability of Richter’s model with the attenuation effects in western Canada should be carefully examined to ensure reliable *M*_{L} estimates at all distances.

Since 2013, several local and regional seismic networks have been deployed in western Alberta and northeastern British Columbia for induced seismicity monitoring. These networks generated a rich ground‐motion dataset of seismic and blast events in the region. One of the earthquakes that was large enough to be recorded by multiple networks over a wide distance range is the 13 June 2015 *M*_{w} 4.0 Fox Creek, Alberta, earthquake. A comparison of *M*_{L} estimates determined based on the standard Richter (1958) relation for the Fox Creek event has uncovered some inconsistencies. As shown in Figure 1, *M*_{L} values obtained at different stations attain an increasing trend with distance, suggesting that Richter’s *M*_{L} formula does not comply with the attenuation effects in WCSB. This results in inconsistent magnitude estimates between local and regional networks, which may introduce inaccuracies into magnitude–recurrence models and GMPEs. Additionally, such discrepancies pose problems in implementation of induced seismicity traffic light protocols, effectively.

## Ground‐Motion Data

Motivated by observations in Figure 1, I investigate the distance‐dependent amplitude decay of ground motions in WCSB. In this regard, high‐quality ground‐motion records of manually reviewed events in the region are compiled from local and regional seismic networks in western Alberta (September 2013–September 2015) and British Columbia (March 2014–April 2016) (see Data and Resources). Events recorded by multiple networks are identified based on their origin times and locations.

An issue which required attention in data preparation is the discrimination of mining explosions picked up by the regional TransAlta network in western Alberta. Figure 2 shows three event clusters collocated with active mines in the region. The lack of *S*‐wave arrivals and the presence of large‐amplitude emergent surface waves in close‐distance seismograms suggest that most of these events are associated with mining operations. The resemblance of surface waves from seismic and blast events makes the visual discrimination of waveforms more difficult at regional distances. Law *et al.* (2015) showed that the average normalized cumulative energy of waveforms obtained at regional distances from suspected blast clusters in western Alberta attain a slower rate of increase with time than that of confirmed earthquake clusters in the region.

Here, event times and their proximity to an operating mine are employed to identify potential blast events. Explosions are generally set off during daytime hours, even if some mines operate 24 hrs. Figure 3 shows the time‐of‐day pattern of events detected in the vicinity of mines in western Alberta (<75 km). It indicates that 98% of the activity near mines occurs during working hours, attaining a peak between 1 p.m. and 2 p.m. local time. The nonuniform distribution of event times is an indicative of human activity. Thus, I classify events that occur between 10 a.m. and 6 p.m. local time, within 75 km of a mine, as suspected blast events. Ninety‐four percent of these events are located less than 30 km distance from a mine, and a small portion of them attain larger offsets primarily due to poor azimuthal coverage and lack of close stations in early stages of the deployment of TransAlta network. Suspected blast events within 30 km radius of a mine are considered in the analysis. It should be acknowledged that there is some subjectivity in this classification, and a few events might be misclassified. However, it will have minimal effect on the derived *M*_{L} relation because radiations from both seismic and blast events show similar attenuation attributes (discussed in the Insights on the Attenuation Attributes in WCSB section).

The compiled ground motions were recorded by three‐component broadband seismometers, most of which have a flat response from 0.05 up to 100 Hz. Ground motions are corrected for instrument response and converted to synthetic WA seismograms in frequency domain. The transfer function of standard WA instrument is computed for a free period of 0.8 s, a damping constant of 0.7, and a static magnification of 2080 (Uhrhammer and Collins, 1990; International Association of Seismology and Physics of the Earth’s Interior [IASPEI], 2013). The amplitude of the largest swing from the *P* arrivals onward are automatically picked on simulated WA seismograms and are then manually verified by analysts (see Data and Resources). The peak amplitudes are measured (in millimeters) as the half peak‐to‐peak amplitude, on each horizontal component (Richter, 1958; IASPEI, 2013). The 96% of the periods associated with the peak WA amplitudes (*T*_{WA}) are between 0.1 s<*T*_{WA}<1 s. In analysis, I exclude ground motions of *T*_{WA}>2 s, to remove data potentially contaminated by low‐frequency noise. I consider events and stations with at least five amplitude measurements up to a distance of 600 km. Based on the above criteria, a total of 105,884 horizontal‐component amplitude readings from 7648 earthquakes are retained. Figure 4 shows the magnitude and distance distribution of the compiled amplitude dataset, in which event magnitudes are computed based on Hutton and Boore (1987) for preliminary assessment. The majority of earthquakes have focal depths shallower than 5 km, attaining an average depth of 2.8 km, which are primarily smaller than *M*_{L} 2.0 and recorded by local networks at distances *R*<30 km. To improve the data distribution at large distances, the amplitude dataset is supplemented with ground motions from suspected blast events, after examining their attenuation attributes (discussed in the Insights on the Attenuation Attributes in WCSB section). Only blast events detected within 30 km radius of a mine are considered to minimize the mapping of location uncertainty into the derived *M*_{L} model. I assume zero depth for mining blasts. The magnitude and distance distribution of an additional 20,512 amplitude readings from 1141 suspected blast events are also presented in Figure 4. The compiled dataset provides a good azimuthal coverage in western Alberta and northeastern British Columbia as shown in Figure 5.

## Insights on the Attenuation Attributes in WCSB

I examine the decay of ground‐motion amplitudes with distance to gain insights on the regional attenuation effects. This is done by scaling WA amplitudes from events of different sizes to a common level at a reference distance. In this regard, two reference distance ranges where both earthquakes and blast events were well recorded at close and far distances are selected: (1) 10–25 km and (2) 100–150 km. For each event, I calculate the geometric mean of amplitudes at a given distance range and normalize individual amplitudes at all distances by the mean value. This exercise is done event‐by‐event, for each reference distance bin separately. It effectively removes the differences in source effects, revealing the attenuation characteristics in the region. It should be noted that the normalized amplitudes still include site effects relative to the average site condition in each reference distance bin.

Figure 6 shows attenuation of normalized WA amplitudes (*A*_{norm}) with distance. By definition, the average *A*_{norm} for all events attain 1 at reference distance bins. A comparison of amplitude decay for earthquakes and blast events suggests that both event types show similar attenuation attributes. Some differences are observed between the two, particularly at distances where empirical data are scarce. This might be due to residual radiation pattern and site effects that are not well sampled over the region. Figure 6 shows that there is a region where amplitude decay slows markedly between 100 and 200 km. This is believed to be due to the Moho‐bounce effect, a consequence of reflected and refracted phases joining the direct waves, modifying the attenuation pattern at intermediate distances. Similar effects were also observed in other regions of North America between ∼70 and ∼150 km, where the exact distance range and strength of Moho‐bounce depends on crustal thickness, velocity structure, dip angle of Moho boundary, and regional anelastic attenuation (Bakun and Joyner, 1984; Burger *et al.*, 1987; Somerville and Yoshimura, 1990; Atkinson and Mereu, 1992; Atkinson, 2004).

*A*_{norm} values in Figure 6 attain apparently slower attenuation rates at distances *R*<3 km, for events well recorded at close distances. It is unlikely to be due to saturation effects because the majority of *A*_{norm} values attaining a constant level at *R*<3 km were obtained from events *M*_{L}<1.0. It could be due to that relatively shallow depth estimates of poorly constrained event solutions are mapping into hypocentral distances at *R*<3 km.

## Model and Regression Analysis

A generalized *M*_{L} relation can be written as (1)(Hutton and Boore, 1987), in which log(*A*) is the logarithm of one‐half peak‐to‐peak WA amplitude (*A*, in millimeters) measured on a horizontal‐component ground motion, at a hypocentral distance *R* (in kilometers); −log*A*_{0} is the regional distance correction model evaluated for the same distance *R*; and *S* is the station correction that accounts for systematic differences in amplitudes between different stations such as local site effects.

The −log*A*_{0} model projects the recorded amplitudes to the reference distance of 100 km. Thus, it should capture the overall attenuation properties of the region of interest to achieve unbiased *M*_{L} estimates with distance. In this study, I delineate −log*A*_{0} for WCSB using a trilinear geometrical spreading model, to account for observed attenuation effects (2)in which the first two terms represent corrections for the region‐specific geometric spreading and anelastic scattering effects, respectively. Equation (2) is constrained to attain 0.001 mm amplitude at 100 km distance (i.e., −log*A*_{0}=3) to maintain Richter’s original definition of zero‐magnitude earthquake. In this respect, both terms are expressed relative to the reference distance, attaining zero (in log units) at 100 km. The geometric‐spreading correction (GS) for WCSB is given as (3)in which *Z* is defined as a hinged trilinear function of distance: (4)The *b*_{1}, *b*_{2}, and *b*_{3} parameters represent rates of geometric spreading at three distance ranges defined by transition distances *R*_{1} and *R*_{2}.

Observed WA amplitudes are regressed based on equations (1)–(4) to determine coefficients of the regional distance correction model (*b*_{1}, *b*_{2}, *b*_{3}, *R*_{1}, *R*_{2}, and *γ*). Also, a correction factor (*S*) for each station that represents the overall discrepancy in WA amplitudes between stations located at different site conditions is computed simultaneously. Site effects are typically determined relative to a known site condition to avoid trade‐offs between source and site terms. In this study, however, site conditions are unknown for the recording stations. Thus, correction factors are computed relative to the average site condition by constraining the *S* terms to attain zero when averaged over all stations. This allows consistent *M*_{L} estimations with reduced overall scatter of predictions across different stations.

Motivated by similar attenuation behaviors observed for earthquakes and blast events, amplitude datasets of the two event types are combined. As presented in Figure 4, most of the earthquake ground motions were obtained from events of *M*_{L}<1.0. To limit the overall weight of small events in analysis, I use all earthquakes with *M*_{L}≥1.0 (1676 events) and an additional 2000 randomly selected local earthquakes with *M*_{L}<1.0. A total of 59,469 amplitudes from earthquakes and 20,512 amplitudes from blast events obtained at 94 stations are used in regressions. Regional coverage of the ray paths for events used in regression analysis is similar to that of the entire catalog (Fig. 5).

## Results and Discussion

I grid‐search transition distances within 50 km≤*R*_{1}≤150 km and 100 km≤*R*_{2}≤300 km ranges with 10 km increments and calculate model coefficients for each *R*_{1}–*R*_{2} combination via regression analysis. Overlapping *R*_{1} and *R*_{2} search windows also consider scenarios of bilinear geometrical spreading with no Moho‐bounce (100 km≤*R*_{1}=*R*_{2}≤150 km). The best‐fitting parameter set is selected by minimizing the mean of absolute residuals. The coefficients of best‐fitting −log*A*_{0} model are listed in Table 1. Station corrections, computed relative to the average of all stations from local and regional networks, are shown for 48 public stations in Table 2. The resultant distance correction model for WCSB is given as (5)

Figure 7 illustrates the −log*A*_{0} model for WCSB (equation 5) in comparison to those of standard *M*_{L} relations that are commonly used in absence of a regional magnitude formula. Distance corrections obtained from observed amplitudes by removing differences in source and site effects based on equation (1) (i.e., −log(*A*_{ij})+*M*_{L,i}−*S*_{j}, in which *M*_{L,i} is the median magnitude of event *i* and *S*_{j} is the correction for station *j*) are also shown in the figure. The WCSB model is in good agreement with the empirical data, attaining a strong Moho‐bounce effect between distances from 100 to 220 km; this is consistent with the observations made in Figure 6. The standard *M*_{L} relations, however, fail to capture the attenuation attributes in WCSB. Both Hutton and Boore (1987) and Eaton (1992) models overcorrect ground‐motion amplitudes for attenuation effects at distances less than 30 km. For the Richter (1958) model, amplitudes are overcorrected at that first 10 km and undercorrected between 10 and 30 km. Discrepancy between the WCSB and standard models increases with decreasing distance. Both WCSB and standard models attain similar values between 30 and 100 km, suggesting that events primarily recorded within this distance range would attain similar magnitudes, regardless of the *M*_{L} relation used. In fact, by definition, all −log*A*_{0} models should attain a value of 3 at 100 km to maintain Richter’s zero‐magnitude earthquake reference, regardless of the attenuation characteristics of the tectonic environment. In Figure 7, the largest discrepancy between the WCSB and standard *M*_{L} relations is observed at distances larger than 100 km. All models attain similar rates at *R*>220 km. However, the Moho‐bounce effect results in a much smaller distance correction in WCSB between 100 and 220 km, causing a shift in −log*A*_{0} values at further distances compared to the standard models.

I further assess the performance of the WCSB *M*_{L} formula in Figure 8, which shows the variation of magnitude residuals as a function of distance for all events in the compiled dataset. The magnitude residuals are defined as *M*_{L,ij}−*M*_{L,i}, in which *M*_{L,ij} is magnitude estimate for station *j* of event *i* based on the WCSB model and *M*_{L,i} is the median magnitude of the event. Figure 8 indicates that average residuals attain values around zero, with no distance dependence up to 600 km. This is valid for both earthquakes and blast events. There is some deviation of residuals from the zero line at *R*<2 km. This is primarily due to inaccurate depth estimations of poorly constrained event locations mapping into the hypocentral distance. Overall, the derived relation well captures the attenuation characteristics in WCSB over a wide distance range, providing reduced scatter of *M*_{L} estimations across different stations due to empirical station corrections (*S*). It is noteworthy that the developed model is specific to WCSB and requires an assessment on the compatibility of −log*A*_{0} term before it is implemented to other regions.

I compare event magnitudes determined based on the standard *M*_{L} relations against those obtained from the WCSB model, in Figure 9. The overcorrection of amplitudes for attenuation effects at close and far distances by standard −log*A*_{0} models result in overestimation of event magnitudes. Specifically, Hutton and Boore (1987) and Eaton (1992) models overestimate earthquake magnitudes, on average, by 0.45 and 0.62 units, respectively. The magnitude discrepancy is smaller for the Richter (1958) model, 0.34 units for earthquakes, because it undercorrects amplitudes at 10 km<*R*<30 km range, which reduces the overall bias in event magnitudes. It should be noted that the error in *M*_{L} estimates differs depending on the distance range at which events are mostly recorded. For instance, the Richter (1958) model attains a larger bias in *M*_{L} estimates, for events primarily recorded by local stations (*M*_{L}<1.0). This is due to the increased discrepancy in the −log*A*_{0} model at close distances. On the contrary, events primarily recorded between 30 and 100 km attain minimal bias in *M*_{L} estimates of standard relations. More far‐distance stations (*R*>100 km) contribute to magnitude calculations with increasing event size. This increases the chance of having biased *M*_{L} estimates from standard magnitude relations.

*M*_{L} is commonly adopted in seismic monitoring applications as a driver of operational protocols designed to mitigate induced seismicity risks. AER (2015) explicitly defines the actions to be taken as response to induced events in western Alberta in terms of staged *M*_{L} thresholds. When an *M*_{L}≥2.0 event is detected during hydrocarbon production, operators should invoke their response plans in a manner that reduces the probability of having further seismic activity. In case of an *M*_{L}≥4.0 event, operators should immediately cease their operations until further notice. Obtaining reliable *M*_{L} estimates is important in terms of effectiveness of traffic light protocols as well as high costs associated with red‐light operation shutdowns. The overestimation of magnitudes by standard *M*_{L} relations may have critical implications in this respect. For instance, Figure 10 shows the largest earthquakes detected by GSC in the Fox Creek region between September 2013 and September 2015. During this time period, 24 events of *M*_{L}≥3.0 were reported based on the Richter (1935) model, two of which exceeded the red‐light threshold (*M*_{L} 4.0). However, *M*_{L} values obtained from the WCSB model suggest that event magnitudes were overestimated by 0.37 units, on average, and only nine of them were actually greater than or equal to *M*_{L} 3.0 (note that besides incompatibility of Richter, 1935, for WCSB, differences in azimuthal coverage of stations used and phases on which peak amplitudes were obtained may also be attributed to the observed discrepancy in Fig. 10).

Inconsistencies in *M*_{L} estimates may limit derivation of effective operational techniques for mitigation of induced seismicity. Additionally, such discrepancies can map into the GMPEs and magnitude recurrence relations, resulting in inaccurate hazard calculations when *M*_{L} estimates are converted to moment magnitude (*M*_{w}). In this regard, the relationship between the two magnitude scales is of particular interest. Recent studies (e.g., Edwards *et al.*, 2010; Bethmann *et al.*, 2011; Ross *et al.*, 2016) show that *M*_{w} attains larger values than *M*_{L} for small earthquakes. In Figure 11, I examine this relationship using well‐recorded events near Fox Creek, Alberta, selected over a wide magnitude range. Regional moment tensor solutions could be obtained for the four largest events only. Therefore, I determine *M*_{w} magnitudes based on spectral fitting technique, in which the apparent displacement source spectrum is matched with the Brune (1970) model at low frequencies. Alternatively, I also calculate *M*_{w} magnitudes from recorded response spectral amplitudes following Atkinson *et al.* (2014). Figure 11 indicates that *M*_{w} values from alternative methods are in good agreement and attain consistently larger values than *M*_{L} for *M*_{w}<3.3. The magnitude discrepancy increases with decreasing event size. These observations are consistent with the *M*_{L}–*M*_{w} relationship proposed by Ross *et al.* (2016), which suggests that accurate modeling of regional attenuation effects for WCSB ensures reliable *M*_{L} estimates, which in turn allows reliable *M*_{w} estimates via magnitude conversion relationships. For *M*_{w}>3.3 events, the two magnitude scales appear to attain similar values although empirical data are limited for drawing a robust conclusion.

## Conclusion

I develop an empirically well‐constrained *M*_{L} relation for earthquakes in WCSB. A comparison of ground‐motion amplitudes from earthquakes and mining blasts in the region shows that both event types decay at similar rates with distance and demonstrate a considerable Moho‐bounce effect between 100 and 220 km. Motivated by this observation, I merge the two amplitude datasets and model the regional attenuation using a trilinear model to account for the observed shape of attenuation. I also determine a site correction for each station, which enables consistent *M*_{L} estimations across different stations, regardless of their site conditions.

The standard *M*_{L} formulas derived from California data do not comply with the attenuation effects in WCSB, overcorrecting amplitudes at close and far distances. This results in overestimated *M*_{L} values by 0.3–0.6 units, depending on the magnitude relation used. Because of the increased inconsistency with regional attenuation attributes at close distances, the standard magnitude relations attain larger bias in *M*_{L} values obtained from local networks. The derived WCSB relation ensures accurate *M*_{L} estimates over a wide distance range (up to 600 km), based on the robust modeling of attenuation effects in the region. A comparison of *M*_{L} and *M*_{w} scales over a wide magnitude range indicates that *M*_{w} attains larger values than *M*_{L} for *M*_{w}<3.3. The discrepancy between the two increases with decreasing magnitude, consistent with the findings of Ross *et al.* (2016). *M*_{L} and *M*_{w} magnitudes show a good agreement for larger events. These results highlight the importance of reliable *M*_{L} calculations, particularly in induced seismicity monitoring and *M*_{w} estimation by magnitude conversion relations.

## Data and Resources

Ground motions of this study are primarily compiled from Nanometrics‐operated local and regional seismic networks in western Alberta (September 2013–September 2015) and British Columbia (March 2014–April 2016). Data from real‐time accessible public stations of Canadian National Seismic Network (CNSN), Canadian Rockies and Alberta Network (CRANE), Regional Alberta Seismic Observatory for Earthquake Studies Network (RAVEN), and U.S. National Seismic Network (USNSN) are also included in the study. The automatic event detections, phase arrivals, event locations, and amplitude picks are manually reviewed by Nanometrics geophysical analysts Andrew Reynen, Caitlin Broderick, and Francis Tong. Regression analyses are performed using MATLAB (www.mathworks.com/products/matlab, last accessed August 2016) and maps are produced by QGIS software.

## Acknowledgments

I would like to thank Gail M. Atkinson, Dario Baturan, Sepideh Karimi, Andrew Law, Kit Chambers, Bill Westwood, and the two anonymous reviewers for helpful discussions and constructive feedback on the results of the study. I would like to acknowledge Canbriam Energy, Canadian Natural Resources Limited, and Repsol Oil & Gas Canada for contributing event catalogs and waveform data obtained from their local networks, which enriched the empirical dataset significantly at close distances. This study was financially supported by Nanometrics Inc.

- Manuscript received 29 August 2016.